Confidence-Gated Cloud-Edge Cascade Triage via Variational Risk Minimization for Medical Imaging

Xinye Yang; Zhusi Zhong; Scott Collins; Michael Bernstein; Grayson Baird; Terrence Healey; Michael Atalay; Mahesh Jayaraman

eess.IVOct 1, 2026

Confidence-Gated Cloud-Edge Cascade Triage via Variational Risk Minimization for Medical Imaging

Authors: Xinye Yang, Zhusi Zhong, Scott Collins, Michael Bernstein, Grayson Baird, Terrence Healey, Michael Atalay, Mahesh Jayaraman,

Organizations: Department of Computer Science, Brown University, Providence, RI, USA · Brown University and Brown University Health, Providence, RI, USA · School of Computing and Information Sciences, Florida International University, Miami, FL, USA

Abstract

Emergency chest X-ray (CXR) triage has a structural modality gap: reports arrive after triage decisions, yet multimodal foundation models require image-text inputs. We present Variational Risk Minimization (VRM), a distillation framework that treats LVLM-generated report variants as Monte Carlo samples of latent clinical interpretations. Rather than distilling from a single teacher target, VRM learns from a variationally marginalized teacher distribution, enabling uncertainty-aware supervision under missing-modality constraints. Under matched encoder families, VRM outperforms direct fine-tuning baselines and improves calibration with strong recovery from hallucinated supervision. Marginalized supervision reduces report-selection instability. In our compact edge-student instantiation, a confidence-gated cascade reaches AUC 0.941 at 103ms average latency with 20.3% cloud escalation, yielding an explicit reliability-latency operating point for cloud-edge clinical workflows.

Figures & tables

Figure 1: VRM Framework: Distilling multimodal foundation models for resource-efficient medical triage. Our approach transfers knowledge from a cloud-based multimodal teacher (BiomedCLIP) to a lightweight, image-only student model by variationally marginalizing LVLM-generated text variations, enabling expert-level performance on consumer-grade hardware without requiring text at inference time.

Figure 2: Architecture of the Teacher Model. It employs a Latent Query Alignment (LQA) mechanism to bridge the modality gap between the image encoder and text encoder using learnable queries, cross-attention, and self-attention layers.

Dataset	Images	Patients	% Urgent	Scan Type
MIMIC-CXR	75,422	13,076	23.4%	AP/PA/Lateral
CheXpert	44,863	13,048	19.7%	AP/PA
IU-CXR	7,470	3,955	16.5%	AP/PA/Lateral
PadChest	32,174	13,518	21.1%	AP/PA/Lateral
RIH-CXR	9,077	9,077	28.7%	AP/PA

Table 1: Dataset Characteristics

Original Label	Urgency Class	Rationale
Pneumothorax	Urgent	Requires immediate intervention
Large Pleural Effusion	Urgent	Potential respiratory compromise
Tension Pneumothorax	Urgent	Life-threatening
Pneumonia (acute)	Urgent	Requires antibiotics within 4 hours
Pulmonary Edema	Urgent	Cardiac/respiratory failure
Cardiomegaly (severe)	Urgent	CTR $>0.6$ , heart failure risk

Table 2: Urgency Label Mapping

Model	Inference Mode	MIMIC AUC	OOD Avg AUC	Avg Latency
Direct FT Foundation Baselines
Standard KD (Logits + Features)	Image (edge)	$0.748\pm 0.007$	$0.708\pm 0.009$	16ms
MC Dropout (EVA-X backbone, 20 stochastic passes)	Image (edge, stochastic)	$0.833\pm 0.007$	$0.794\pm 0.009$	412ms
DINOv2-RadBench Direct FT ( Baharoon et al., 2023 )	Image (edge)	$0.821\pm 0.006$	$0.780\pm 0.008$	24ms
CXR Foundation Direct FT ( Kiraly et al., 2024 )	Image (edge)	$0.824\pm 0.006$	$0.784\pm 0.008$	22ms
EVA-X Direct FT ( Yao et al., 2025 )	Image (edge)	$0.828\pm 0.006$	$0.788\pm 0.008$	29ms

Table 4: External Benchmark Reference with Contemporary LVLMs. Direct FT foundation baselines and a classical uncertainty baseline (MC Dropout) are compared against VRM with encoder transfer, while open-source LVLM direct inference (zero/few-shot) is reported as deployment context.

Config	MIMIC	OOD Avg	Escalation Rate	Avg Latency
VRM Student	0.842	0.806	0.0%	14ms
Teacher	0.962	0.915	100.0%	450ms
Cascade	0.941	0.892	20.3%	103ms

Table 8: Cascade Triage Performance-Delay Trade-off. Escalation rate indicates fraction of cases routed to cloud teacher.

Figure 3: VRM Student AUC results. Effect of Monte Carlo Samples (

K

) on Supervision Quality and Variance. Increasing

K

improves discrimination on both ID (MIMIC) and OOD (RIH-CXR) while reducing teacher prediction variance, consistent with the marginalization hypothesis. Here, teacher prediction variance is defined as

\mathbb{E}_{x}\!\left[\mathrm{Var}_{k=1}^{K}\!\left(T(x,t_{k})\right)\right]

Figure 4: Hallucination Recovery Analysis. (a) Prediction error rate on cases with at least one hallucinated report decreases progressively with

K

. (b) Recovery rate reaches

\sim

66% at

K=5

and

\sim

80% at

K=10

, demonstrating VRM’s meaningful noise tolerance through variational marginalization.

Figure 5: Individual Report vs. Marginalized Ensemble. (a) Marginalization (0.842) exceeds the average individual report (0.808) and the best single report (0.832), approaching the Oracle per-case bound (0.856). (b) Per-sample prediction standard deviation is reduced by 53% (0.095

\to

0.045), indicating more stable supervisory targets.

Figure 6: Report Selection Stability Bootstrap Test (

B=100

). (a) AUC distributions show

K=1

(red) spans

\sim

10% range while

K=5

(blue) is tightly concentrated (

\sigma=0.005

). (b) Running-mean AUC confirms rapid convergence for

K=5

vs. sustained variability for

K=1

, demonstrating reduced report-selection risk under marginalization.

LVLM Sampler	Params	Domain	MIMIC AUC	OOD Avg	BERTScore
InternVL2-26B ( Chen et al., 2024b )	26B	Gen	0.779	0.752	0.571
Qwen2.5-VL-32B ( Wang et al., 2024 )	32B	Gen	0.786	0.758	0.603
HuatuoGPT-Vision-34B ( Chen et al., 2024a )	34B	Med	0.839	0.804	0.722
MedGemma-27B (default)	27B	Med	0.842	0.806	0.731
References: Standard KD: 0.748 Best direct FT (EVA-X): 0.828

Table 13: LVLM Sampler Sensitivity. All models use

K=5

and identical VRM training. Domain: medical (Med) vs. general-purpose (Gen). BERTScore is against reference radiology reports.

Knowledge Distillation

Knowledge distillation transfers useful behavior or representations from larger or specialized models into more compact or adaptable learners.

See all 821 papers

Reliability Stress Tests and Decision-Time Routing for Chest X-ray Vision-Language Models
Oct 1, 2026Xinye Yang, Zhusi Zhong, Scott Collins +3Medical VLMs VLM Evaluation
Department of Computer Science Brown University Providence, RI, USA · Brown University Brown University Health Providence, RI, USA · School of Computing and Information Sciences Florida International University Miami, FL, USA
Adapt or Calibrate? How Local Labels Buy Certified Coverage in Medical VQA
Jun 14, 2026Reza Khanmohammadi, Kundan Thind, Mohammad M. GhassemiVLM Evaluation Confidence Estimation for VLMs
Michigan State University East Lansing, MI, USA · Henry Ford Health Detroit, MI, USA
CXRMate-2: Structured Multimodal Temporal Embeddings and Tractable Reinforcement Learning for Clinically Acceptable Chest X-ray Radiology Report Generation
Apr 21, 2026Aaron Nicolson, Elizabeth J. Cooper, Hwan-Jin Yoon +7Radiology Report Generation Radiology VLMs
Australian e-Health Research Centre, CSIRO Health and Biosecurity, Brisbane & Melbourne, Australia · Princess Alexandra Hospital, Metro South Health, Brisbane, Australia · Queensland Digital Health Centre, University of Queensland, Brisbane, Australia +2

Oct 1, 2026J/K move · Enter open · S save

Reliability Stress Tests and Decision-Time Routing for Chest X-ray Vision-Language Models

Xinye Yang, Zhusi Zhong, Scott Collins, Grayson Baird+2

Department of Computer Science Brown University Providence, RI, USA · Brown University Brown University Health Providence, RI, USA · School of Computing and Information Sciences Florida International University Miami, FL, USA

Medical VLMs VLM Evaluation Radiology VLMs Chest X-Ray Classification

Open paper AskPDF arXiv

Medical vision-language model (VLM) evaluation is sensitive to workflow design, prompting strategy, and benchmark construction, yet most studies treat these factors in isolation. We introduce a reliability stress test for chest X-ray interpretation built on two balanced datasets (a private report-backed set and a curated MIMIC subset). Three medical VLMs (CheXagent, MedGemma-4B, and MedGemma-27B) are evaluated across three prompt styles and two workflows (single-VLM and multi-agent), producing 36 configurations. We show that exact-match accuracy alone can overstate the effectiveness of conservative models that default to "Normal" predictions. Diagnostic reliability also depends heavily on model family and scale: multi-agent reasoning helps some configurations but hurts others. Building on these observations, we propose a decision-time routing framework that selectively escalates to multi-agent inference only when beneficial, improving the cost-quality trade-off over fixed workflows. Our results highlight the need for evaluation protocols that jointly consider prompt sensitivity, failure-mode diversity, and workflow choice before clinical deployment.

Abstract

Knowledge Distillation

Explore similar work

Reliability Stress Tests and Decision-Time Routing for Chest X-ray Vision-Language Models

Adapt or Calibrate? How Local Labels Buy Certified Coverage in Medical VQA

CXRMate-2: Structured Multimodal Temporal Embeddings and Tractable Reinforcement Learning for Clinically Acceptable Chest X-ray Radiology Report Generation